AI Glossary

A mathematical function applied to neuron outputs that introduces non-linearity, enabling networks to learn complex patterns.

Inputs deliberately designed to cause a model to make mistakes, used for testing robustness and security.

An AI system that can perceive its environment, make decisions, and take actions to achieve specific goals autonomously.

The process of ensuring AI systems behave in accordance with human values and intentions, crucial for safety.

A technique that allows models to focus on specific parts of the input when processing information, crucial for transformer architectures.

A neural network that learns to compress data into a lower-dimensional representation and reconstruct it.

A model that generates outputs sequentially, where each new token depends on previously generated tokens.

An algorithm for training neural networks by computing gradients of the loss function with respect to weights.

The number of training examples processed together in one iteration, affecting training speed and memory usage.

A decoding strategy that explores multiple possible sequences simultaneously to find high-quality outputs.

Bidirectional Encoder Representations from Transformers - a pre-training method for NLP that reads text bidirectionally.

Systematic errors in model predictions, often reflecting biases in training data or model architecture.

A metric for evaluating machine translation quality by comparing generated text to reference translations.

When a model loses previously learned knowledge while learning new information, a challenge in continual learning.

A prompting technique that encourages models to show step-by-step reasoning, improving complex problem-solving.

A saved snapshot of model weights during training, allowing resumption or evaluation at specific points.

Contrastive Language-Image Pre-training - a model that learns visual concepts from natural language supervision.

The maximum number of tokens a model can process at once. Larger context windows allow understanding longer documents.

A training approach that learns representations by contrasting similar and dissimilar examples.

A neural network architecture designed for processing grid-like data such as images, using convolutional layers.

An attention mechanism where queries from one sequence attend to keys and values from another sequence.

Techniques for artificially expanding training datasets by creating modified versions of existing examples.

The component of a model that generates output sequences, often used in transformer architectures.

A generative model that learns to reverse a gradual noising process, used in image generation like Stable Diffusion.

Training a smaller 'student' model to mimic a larger 'teacher' model, creating efficient compressed models.

A regularization technique that randomly deactivates neurons during training to prevent overfitting.

Vector representations of text, images, or other data that capture semantic meaning in a continuous space.

The component of a model that processes input sequences into representations, often used in transformers.

One complete pass through the entire training dataset during the training process.

A model's ability to learn new tasks from just a few examples, demonstrated through in-context learning.

The process of adapting a pre-trained model to a specific task or domain by training on specialized data.

Floating Point Operations Per Second - a measure of computational performance for AI hardware.

Large-scale models trained on broad data that can be adapted to many downstream tasks through fine-tuning.

Generative Adversarial Network - two neural networks competing to generate realistic synthetic data.

GPT-Generated Unified Format - a file format for storing quantized language models efficiently.

Generative Pre-trained Transformer - an autoregressive language model architecture that generates text left-to-right.

An optimization algorithm that iteratively adjusts model parameters to minimize the loss function.

A simple generation strategy that always selects the most probable next token at each step.

When an AI model generates false or nonsensical information presented as fact, a common challenge in LLMs.

Configuration settings for training (like learning rate) that are set before training begins, not learned.

A model's ability to learn from examples provided in the prompt without updating its parameters.

The process of using a trained model to make predictions or generate outputs on new data.

Fine-tuning models on instruction-following datasets to improve their ability to follow user commands.

Techniques to bypass safety guardrails in AI models, a security concern for deployed systems.

See Distillation - transferring knowledge from a large model to a smaller, more efficient one.

A comprehensive benchmarking framework that evaluates AI models across multiple dimensions beyond traditional metrics.

The time delay between input and output in AI systems, critical for real-time applications.

A technique that normalizes activations across features to stabilize and accelerate training.

A hyperparameter controlling how much model weights are adjusted during training, crucial for convergence.

Large Language Model Meta AI - Meta's family of open-source language models ranging from 7B to 405B parameters.

Large Language Model - neural networks with billions of parameters trained on vast text corpora.

Low-Rank Adaptation - an efficient fine-tuning method that updates only a small subset of model parameters.

A mathematical function measuring the difference between model predictions and actual values, guiding training.

Long Short-Term Memory - a recurrent neural network architecture designed to handle long-term dependencies.

A pre-training objective where models predict masked tokens in text, used in BERT-style models.

An architecture where multiple specialized sub-models (experts) are combined, with routing determining which experts process each input.

Techniques like quantization and pruning to reduce model size and computational requirements.

An attention mechanism with multiple parallel attention layers, allowing models to focus on different aspects.

Models that can process and generate multiple types of data (text, images, audio, video) in a unified framework.

Automated methods for discovering optimal neural network architectures for specific tasks.

Techniques for scaling data or activations to improve training stability and convergence speed.

A generation strategy that samples from the smallest set of tokens whose cumulative probability exceeds a threshold.

Learning to perform a task from a single example, a special case of few-shot learning.

When a model learns training data too well, including noise, resulting in poor generalization to new data.

The learnable weights in a neural network. More parameters generally mean more capacity but require more compute.

A metric measuring how well a language model predicts text, with lower values indicating better performance.

Information added to embeddings to represent token positions in sequences, crucial for transformers.

Initial training phase on large datasets to learn general representations before task-specific fine-tuning.

The practice of crafting effective input prompts to elicit desired outputs from language models.

Removing unnecessary weights or neurons from a model to reduce size and improve efficiency.

Reducing the precision of model weights (e.g., from 32-bit to 8-bit) to decrease memory usage and increase speed.

Retrieval-Augmented Generation - combining language models with external knowledge retrieval for more accurate responses.

A neural network architecture designed for sequential data, processing inputs one step at a time.

Training approach using human preferences to align model behavior with desired outcomes.

Skip connections that add layer inputs to outputs, enabling training of very deep networks.

Metrics for evaluating text summarization by comparing generated summaries to reference summaries.

Methods for selecting tokens during generation, balancing between randomness and quality.

A mechanism where each position in a sequence attends to all positions, enabling parallel processing in transformers.

Search based on meaning rather than keywords, typically using embeddings and vector similarity.

Models that transform input sequences into output sequences, used for translation and summarization.

An activation function that converts logits into probability distributions over classes.

An open-source text-to-image diffusion model capable of generating high-quality images from text descriptions.

Training approach using labeled data where the model learns to map inputs to known outputs.

A parameter controlling randomness in generation. Higher values increase creativity, lower values increase determinism.

A multi-dimensional array used to represent data and parameters in neural networks.

The basic unit of text processing in language models, typically representing words or subwords.

The process of breaking text into smaller units (tokens) that models can process, typically words or subwords.

A generation strategy that samples from the k most probable next tokens at each step.

See Nucleus Sampling - samples from tokens whose cumulative probability exceeds p.

Applying knowledge learned from one task to improve performance on a different but related task.

A neural network architecture based on self-attention mechanisms, foundational to modern LLMs and vision models.

When a model is too simple to capture patterns in data, resulting in poor performance on both training and test data.

Training approach using unlabeled data where models discover patterns without explicit supervision.

A subset of data used to evaluate model performance during training and tune hyperparameters.

A generative model that learns probabilistic latent representations of data.

A database optimized for storing and querying high-dimensional vectors, essential for semantic search and RAG.

A transformer architecture adapted for computer vision tasks, treating image patches as tokens.

Gradually increasing the learning rate at the start of training to stabilize optimization.

A regularization technique that penalizes large weights to prevent overfitting.

OpenAI's open-source speech recognition model trained on 680,000 hours of multilingual audio data.

A model's ability to perform tasks it wasn't explicitly trained on, using only natural language instructions.